Boosting Robustness of Deep Neural Networks against Sparsity-aware Adversarial Attacks

Over the past few years, deep neural networks (DNNs) have been used to solve a wide range of real-life problems. However, DNNs are vulnerable to adversarial attacks where carefully crafted input perturbations can mislead a well-trained DNN to produce false results. As DNNs are being deployed into security-sensitive applications such as autonomous driving, adversarial attacks may lead to catastrophic consequences. In this research project, we focus on those adversarial attacks that target energy-efficiency of DNNs. In particular, there are a group of attacks that insert perturbation into inputs of neural networks so that the number of zero values reduces. This will increase energy consumption and latency of machine learning applications. We propose a solution to monitor outputs of neurons and detect these malicious attacks dynamically and in runtime. This project enables TrojAI to offer machine learning algorithms that are protected against sparsity-aware adversarial attacks.

Faculty Supervisor:

Ehsan Atoofian

Student:

Partner:

TrojAI

Discipline:

Engineering

Sector:

Professional, scientific and technical services

University:

Lakehead University

Program: